ESG-Kor: A Korean Dataset for ESG-related Information Extraction and Practical Use Cases

1 Introduction

• Specialized dataset construction requires relevant experts, while English-centered resources and translated datasets have limited applicability or performance in specialized non-English tasks.The paper notes that multilingual pretrained models perform less well than single-language models.Source: §1 · p. 1
• ESG-Kor contains 118,946 manually labeled sentences from sustainability reports, with labels based on objective ESG evaluation-provider rules.Three annotators with relevant knowledge performed the labeling.Source: §1 · p. 2
• Fine-tuned pretrained models appropriately performed ESG sentence extraction and classification, supporting ESG-Kor’s usefulness.The paper also presents use cases extending ESG analysis to SMEs and ESG key issues.Source: §1 · p. 2
• The paper presents SME expansion and ESG key-issue analysis as practical use cases for the dataset and model.The key-issue use case also supports in-depth corporate ESG analysis.Source: §1 · p. 2
• ESG-Kor is described as the only Korean-based professional NLP benchmark dataset manually labeled for ESG.The paper proposes its construction process as a benchmark for ESG datasets in other languages.Source: §1 · p. 2

2.1 Specialized Domain & Minority Language

• Specialized-domain resources include biomedical, legal, geographic, and finance datasets, built for tasks such as QA, evaluation, contract review, and geographic understanding.Examples include BLURB, LexGLUE, CUAD, GeoGLUE, and MultiHIERTT.Source: §2.1 · p. 2
• Language-specific resources span Korean, Japanese, Russian, Spanish, and Arabic datasets, while CMU-MOSEAS provides broader multilingual coverage.The cited resources include medical, QA, clinical, reading-comprehension, and sentiment datasets.Source: §2.1 · pp. 2–3

2.2 ESG Research using Machine Learning

• ESG machine-learning research covers financial indicators, investment portfolios, ratings, sustainability classification, media-based indices, and ESG-related language.The passage identifies limited expertise and relevant ESG datasets as constraints on this research.Source: §2.2 · pp. 3–4

3.1 Sustainability Reports

• Reports from 53 sectors published between 2002 and 2021 were collected, with five higher-data sectors selected and POSCO steel and NAVER IT reports added for utility evaluation.The five selected sectors were automobile, aviation, electronics, finance, and heavy industry.Source: §3.1 · p. 4

3.2 Preprocessing

• PDF text was sentence-split, normalized, spell-checked, deduplicated, and manually reviewed for segmentation or deletion when noisy or unsuitable in length.The reports contained text, tables, and images, but only text was extracted.Source: §3.2 · p. 4

3.3 Labeling

• Three knowledgeable annotators used criteria from ESG rating providers and Korean evaluation agencies, discussing ambiguous sentences and deleting cases without unanimous agreement.Sentences deleted because annotators disagreed were less than 0.1% of the dataset.Source: §3.3 · pp. 4–5

3.4 Exploratory Data Analysis

• 118,946 sentences make up ESG-Kor, which also records issue year and issuing company.Figure 2 shows labeled-class distributions by sector and for the full dataset.Source: §3.4 · p. 5, Figure 2 · p. 5
• Finance had a higher proportion of S labels and fewer E labels than G labels, unlike the broadly similar class distributions across most sectors.The authors interpret this pattern as consistent with finance’s relatively lower environmental impact and greater emphasis on social activities.Source: §3.4 · p. 5, Figure 2 · p. 5
• Sentence lengths ranged from 3 to 233 tokens, averaging 42 with a standard deviation of 20; all sentences fit within the 512-token-or-higher limits used by most models.Very short financial-item sentences were labeled Irrelevant; sentences judged too short or long were manually reviewed and removed.Source: §3.4 · pp. 5–6, Figure 3 · p. 5
• Table 3 presents sentence counts by class and sector.The passage identifies the table’s scope but does not provide its individual counts.Source: §3.3 · p. 5, Table 3 · p. 6

4.1 Experimental Settings

• KoBERT, DistilKoBERT, KB-ALBERT, and KLUE-RoBERTa were fine-tuned using BERT-based models and a cross-entropy objective.The sentence representation was the [CLS] token, combined with a classification layer to estimate class probabilities.Source: §4.1 · p. 6
• The data were split 70%/15%/15% for training, validation, and testing, with batch size 64 and maximum sequence length 256.AdamW was used; the four models had learning rates of 1e-04, 5e-04, 5e-05, and 5e-05, respectively.Source: §4.1 · p. 6

4.2 Classification Result

• The dataset supports multiclass E/S/G classification and binary ESG-relevance classification, but IT-sector-specific experiments were omitted because of limited collected data.IT data remained included in the overall performance evaluation.Source: §4.2 · p. 6
• 85.87% accuracy was highest for KLUE-RoBERTa-large in four-class classification, while DistilKoBERT recorded 84.19%; KLUE-RoBERTa-large reached 86.67% in binary classification.Four-class classification excluded R, which combines ESG information; binary classification grouped E, S, G, and R as relevant.Source: §4.2 · p. 7
• Models with more parameters performed better overall, although KB-ALBERT matched or exceeded KoBERT and DistilKoBERT despite having fewer parameters.The authors suggest KB-ALBERT’s financial-document pretraining may help it understand ESG documents.Source: §4.2 · p. 7

5.1 SMEs ESG Information Classification

• The SME validation used 2022 reports from three companies because SME ESG reports are limited, and the 403-sentence dataset was set up for binary ESG-relevance classification.The same preprocessing and labeling process as ESG-Kor was used.Source: §5.1 · p. 7
• A model trained with ESG-Kor achieved 86.72% accuracy in classifying ESG reports from SMEs, supporting the applicability of large-corporation-trained models in SME settings.Source: §5.1 · p. 7

5 Use Cases

• The use cases extend ESG-Kor to SME classification and key-issue analysis, with key issues organized into 10 main and 33 detailed categories.The issue-based approach targets the difficulty of extracting relevant information from large volumes of company data.Source: §5.2 · pp. 7–8, §5 · p. 7

5.2 Extraction of ESG Key Issues

• The extraction pipeline filters ESG-related sentences, embeds them with key issues, and retrieves the most similar sentences by cosine similarity.For Hyundai Construction reports from 2010–2020, the model extracted 4,588 ESG-related sentences from 6,403; results were compiled into top-ten sets for main issues while including at least one sentence per detailed issue.Source: §5.2 · p. 8
• Across all evaluations on a five-point inclusion scale, extracted sentences averaged 4.67 for all ESG key issues; the highest-scoring category reached 4.9, while Governance Behavior scored lowest at 4.4.The scale ran from 1 (Not included) to 5 (Fully included).Source: §5.2 · p. 8, Table 6 · p. 8

5.3 Discussion

• The discussion identifies potential uses in sustainability-report preparation, investment decisions, company supervision, policy development, and ESG scoring.It also notes that rating agencies use different criteria and may not disclose precise evaluation rules.Source: §5.3 · pp. 8–9

6 Conclusions

• 85.87% multiclass accuracy and 86.67% binary accuracy were recorded by KLUE-RoBERTa-large.The conclusion describes these results as evidence that the model appropriately learned and performed the task.Source: §6 · p. 9
• ESG-Kor contains 118,946 manually labeled sustainability-report sentences across seven Korean industry sectors.The labels followed ESG criteria, and three annotators with relevant knowledge labeled the dataset.Source: §6 · p. 9
• The authors report 86.72% accuracy for the SME evaluation and human-evaluation validation of key-issue extraction.These use cases extend the approach beyond the core dataset construction and classification experiments.Source: §6 · p. 9
• The authors propose the ESG-Kor construction process as a benchmark for building ESG datasets in other countries with minority-language settings.They note that specialized-domain translation is challenging and its performance is limited.Source: §6 · p. 9

Limitation

• ESG-Kor uses sustainability-report text only, leaving tables and graphs outside the dataset’s current scope.The authors plan to broaden the dataset with these other information types.Source: Limitation · p. 9
• The study uses BERT models rather than large language models, citing resource accessibility and potential corporate data-privacy concerns.The authors describe these as considerations in model selection.Source: Limitation · p. 9
• The current work extracts and classifies ESG information but does not yet assign ESG-level scores.The authors identify scoring as a possible future extension using extracted information.Source: Limitation · pp. 9–10

Ethics Statement

• The dataset contains no socially controversial or ethically inappropriate data, according to the authors.They report no concerns related to the study’s ethics statement.Source: Ethics Statement · p. 10

• Table A.1 reports each model’s parameter count and elapsed time for the 4-class and 2-class classification tasks.Source: §A.1 · p. 13, Table A.1 · p. 13

A.2 MSCI’s ESG key issues

• MSCI’s ESG key issues are organized by environmental, social, and governance domains, then related issues and detailed issues.Source: §A.2 · p. 13, Table A.2 · p. 13

A.3 Human Evaluation

• Ten participants rated the presence of detailed ESG-key-issue items in 10 extracted pieces per domain on a 1-to-5 scale.Source: §A.3 · p. 13
• Corporate Behavior received the lowest human-evaluation score, 4.4, which the authors attribute as likely related to abstract business-ethics content.The report’s business-ethics information was not detailed, although extracted sentences included some ethical-management content.Source: §A.3 · pp. 13–14
• The authors suggest the evaluation result can be extended to developing future ESG scoring systems.Source: §A.3 · p. 14